Completely agreed. For some of these large language models, it would take a long time before inference spend dominates training spend.